Cartoon Spider Walking a Dog in the Rain and Other AI Generated Images

[jack action]

This seems like a relevant update to the thread

https://www.cnn.com/2022/09/03/tech/ai-art-fair-winner-controversy/index.html

This is pretty much the same thing as when photography was giving anyone the power of doing a portrait. That's when artists redefined themselves with abstract art as they weren't needed as much anymore.

 

[Greg Bernhardt]

I've been able to play with Dalle-2 and Midjouney. They are pretty incredible. Still, some work to do to get better and it takes a lot of skill to get what you want from the prompts sometimes, but this is the future.

 

[anorlunda]

What fun! Thank you @Drakkith .

My first try was "micro bikinis". That produced impressive results that I won't post here.

My second try was "induction motor". That produced boring photos of real motors.

My third try was "neural network". That gave me this:

 

[Wrichik Basu]

"Particle accelerator"

"Transverse beam dynamics"

"Quadrupole magnets"

"Particle collision"

 

[Borg]

John Oliver and his cabbage.

 

[DrJohn]

As the uk conservative party has elected Liz Truss as party leader and thus british PM, I tried "Truss wins election". Very realistic, generation images from many people's nightmares

 

[Swamp Thing]

As the uk conservative party has elected Liz Truss as party leader and thus british PM, I tried "Truss wins election". Very realistic, generation images from many people's nightmares

Attachments:

 

[DrJohn]

I thought I'd test it a bit more and used stealth fighter aircraft, dark sky, nighttime.
The results were astonishly realistic as you can see. Then I tried adding the moon. I am not kidding, these are what I got.

 

[Swamp Thing]

https://nymag.com/intelligencer/2022/09/ai-art-is-here-and-the-world-is-already-different.html

 

[jack action]

 

[DennisN]

I recently saw this thread (AI enhanced image of M87 black hole) (by @pinball1970) and started wondering how well AI software would be able to render celestial bodies.

The online AI tools I have tried are no match for the "real" stuff, so I asked one of my friends who has got AI software if he would like to try rendering some space related images.

He did, and I really, really liked the results .
He said he struggled to get the Moon round (as can be seen in some of the pictures) and also that he thinks the software hasn't been trained so much on planetary/space stuff. The software he used is Stablediffusion "with extra training to make more realistic people".

Nevertheless, here are some of the results (post 1/2, for more pictures, see next post):

An astronaut party in space:

First AI man on the Moon?

The Moon and Earth:

The Moon:

Astronauts and landers on the Moon:

Lunar surface:

 

[DennisN]

...continued... post 2/2:

A girl and the Moon:

Girl with a Moon suit:

Two astronauts:

Two astronauts and the Milky Way:

An astronaut on another world:

Girl astronaut and the Milky Way:

Girl astronaut on another world:

Astronaut gear on another world:

Well, that's it from the world of AI and space for this time .

 

[pinball1970]

...continued... post 2/2:

A girl and the Moon:
View attachment 324981

Girl with a Moon suit:
View attachment 324982

Two astronauts:
View attachment 324983

Two astronauts and the Milky Way:
View attachment 324984

An astronaut on another world:
View attachment 324986

Girl astronaut and the Milky Way:
View attachment 324987

Girl astronaut on another world:
View attachment 324988

Astronaut gear on another world:
View attachment 324989

Well, that's it from the world of AI and space for this time .

Some CGI stuff now looks so life like one has to have a double take, GTPchat is on 4 now? Starting to answer like a person and AI is problem solving like a human.

These images are incredible.

 

[SparkBird]

...continued... post 2/2:

A girl and the Moon:
View attachment 324981

Girl with a Moon suit:
View attachment 324982

Two astronauts:
View attachment 324983

Two astronauts and the Milky Way:
View attachment 324984

An astronaut on another world:
View attachment 324986

Girl astronaut and the Milky Way:
View attachment 324987

Girl astronaut on another world:
View attachment 324988

Astronaut gear on another world:
View attachment 324989

Well, that's it from the world of AI and space for this time .

Great pictures! Impressive! Wondering, is there a way to set prompts in such a way as to get a "real photo effect". For example, so that in the airless space on other planets there were no pictures with an open spacesuit. Have there been any attempts to make it happen?

 

[pxl]

Great pictures! Impressive! Wondering, is there a way to set prompts in such a way as to get a "real photo effect". For example, so that in the airless space on other planets there were no pictures with an open spacesuit. Have there been any attempts to make it happen?

Thanks :)

( I made the shots above )

Yeah it's definitely possible - basically prompt ( describe ) the image in a way that would close up the helmet. You can also do inpainting, basically mask the general area of the helmet and then write a prompt specifically for that area ( it will adapt what you prompted to fit the image )

The models are evolving really fast and what you see above is already dated. New models arrive almost daily.

 

[berkeman]

The online AI tools I have tried are no match for the "real" stuff, so I asked one of my friends who has got AI software if he would like to try rendering some space related images.

Great pictures! Impressive! Wondering, is there a way to set prompts in such a way as to get a "real photo effect". For example, so that in the airless space on other planets there were no pictures with an open spacesuit. Have there been any attempts to make it happen?

Thanks :)

( I made the shots above )

Welcome to PF! So you made the images that @DennisN posted? Very impressive. Can you say more about what tools you used to generate them?

 

[pxl]

Thanks nice to be here :)

Sure - happy to help out !!

You need a web UI for StableDiffusion, a few special models and perhaps a Lora or two.

You can run this locally ( if you got the hardware that can deal with the compute ) or you can hire compute in the cloud with places like runpod.io... ( links below )

..here's what you need... ( in the order you likely will need it )

Web UI
https://github.com/AUTOMATIC1111/stable-diffusion-webui

Tutorial to install on MAC...


..and on Windows


Then find models on huggingface or here ( beware...a lot of NSFW...just human nature ) on each image in the gallery you can see exactly what prompt and settings that where used...you can copy them, paste in the first field and then click the arrow and it will fill in all the fields...there's a lot of cool images here and you can turn off the NSFW in settings...highly recommended site...all the latest stuff you find here https://civitai.com/

For the ultra realistic stuff - the images I posted here - I used...
https://civitai.com/models/4201/realistic-vision-v20

How to get super high resolution...


You can accelerate this in the cloud ( 50cent-ish an hour ) with runpod...

Cloud rentals...
https://www.runpod.io/

Tutorial to do this...


I highly encourage to check out these excellent youtubers...they taught me all I know...

In particular...
https://www.youtube.com/@sebastiankamph/videos

..and also...
https://www.youtube.com/@Aitrepreneur/videos
https://www.youtube.com/@SECourses/videos

Good luck :)

//O.

 

[pxl]

Two new shots...you can see the quality is getting better by the week :)

 

[berkeman]

Wow!

 

[DennisN]

Two new shots...you can see the quality is getting better by the week :)

If I understood correctly you can train the software for particular tasks by using other images, or?
For instance, if you would like a more realistic lighthouse you could feed the software with photos of lighthouses?

Well, I've got a bunch of such photos in hi-res, so let me know if you need them .

 

[pxl]

Very kind - thank you !!

I haven't yet trained on photos aside from training a "lora" on my own face just to learn...and from that I realised I need to learn a lot more before I train full models. ( a Lora is a small network trained on a specific subject usually 10 or 15 images that can then be used in combination with a larger model trained on tens of thousands of annotated images ) But you're correct, you could make a Lora using shots of a specific subject like a lighthouse, then remix the look and use of that by combining it with a larger model...that would add a beach and ocean, change the colour and design etc.

I made a few half decent shots of characters that look like me ( easiest way to get a photograph of myself in a suit haha ) but it was overcooked or I did something wrong somewhere and that makes it hard to work with I.E I have a lot more to learn.

There's a strong community that train on all sorts of content - I can imagine not all of it sourced in a responsible way - hence being wary that the output may not be viable for commercial use.

I used to work in advertising so have an interest, now moved on to other things so this is merely for fun.

Another way of doing it is to throw together photos you have taken into a "very messy and clunky composition" then upscale that with a model that will fix all shadows, edges, etc and "just make it look right" based on being trained on "what is looking right" - adding a prompt will make the result even better.

I've done a bunch of that and it's like having a retouch assistant.

Also fun to just give it a poem and see what it makes of it :)

 

[pxl]

..or make strange places and architecture ?

..it's like visiting places simply by describing them :)

 

[pxl]

You can tell I like dystopian abandoned ( ? ) structures :)

 

[Bandersnatch]

Other than the middle one, the structures are clearly not abandoned, as somebody's picking leaves out of the pools and trimming the hedges. Or something is.
Was the prompt 'show me the world a few years hence, when all humans have been made redundant and the AI reigns alone over the depopulated Earth'?

Thanks for sharing the process, btw. The birth(transformation?) of what it means to be an artist is equally fascinating to observe as the advances in the technology themselves.

 

[DaveC426913]

I find myself continually asking how it is doing this?

I mean, I assume when you ask to for a pic of 'a girl in a spacesuit', it looks at common elements of hundreds or thousands of pics that have similar ... tags?
(Anthropomorphized:
"OK I've got 567 images all labelled 'spacesuit', all with these common elements. If I draw a pic that matches those elements, it ought to be a spacesuit." Yes?)

Does it steal stuff, or is it all original? If we look at its input for 'spacesuit', will we find one or a few that are recognizably the inspiration?

 

[pxl]

@Bandersnatch haha this is what I really like about images of "strange places"...inferring their nature from logic like you just did...that immediately builds a story within that location or universe :)

@DaveC426913 here's a video that explains it to some detail...there's a lot more to it but these are the principles...

 

[Algr]

..it's like visiting places simply by describing them :)

That is what the characters in Myst were doing. They would write these books, and it wasn't clear if they were creating worlds or contacting them.

Does it steal stuff, or is it all original?

There are lots of upset artists accusing it of stealing. But if the AI were really doing that, we'd be seeing more examples of people comparing AI output to the "original" images.

 

[SparkBird]

Very kind - thank you !!

I haven't yet trained on photos aside from training a "lora" on my own face just to learn...and from that I realised I need to learn a lot more before I train full models. ( a Lora is a small network trained on a specific subject usually 10 or 15 images that can then be used in combination with a larger model trained on tens of thousands of annotated images ) But you're correct, you could make a Lora using shots of a specific subject like a lighthouse, then remix the look and use of that by combining it with a larger model...that would add a beach and ocean, change the colour and design etc.

I made a few half decent shots of characters that look like me ( easiest way to get a photograph of myself in a suit haha ) but it was overcooked or I did something wrong somewhere and that makes it hard to work with I.E I have a lot more to learn.

There's a strong community that train on all sorts of content - I can imagine not all of it sourced in a responsible way - hence being wary that the output may not be viable for commercial use.

I used to work in advertising so have an interest, now moved on to other things so this is merely for fun.

Another way of doing it is to throw together photos you have taken into a "very messy and clunky composition" then upscale that with a model that will fix all shadows, edges, etc and "just make it look right" based on being trained on "what is looking right" - adding a prompt will make the result even better.

I've done a bunch of that and it's like having a retouch assistant.

Also fun to just give it a poem and see what it makes of it :)

I've seen abt 1 month ago as some guys tested one prompt-poem worked out by different AI tools, generating images and video of the poem text. There were DALL-E, StableDiffusion and 2 another (can't recall names not to find that video). I'd say it was creepy and they did it for a fun. Although, they have shown also retrospective of the same prompt-poem with the same AI tools but made of 3 months before. And that was awful. Than they've shown the same poem with added text-prompts "cleanning out" all the "AI overinvented fragments", setting up parameters of shades, style, creativeness level +destinguishing fragments with added sampling templates etc. The results were really wonderfull. The speed of AI leraning is extreemly high, as you noticed and the level of mastering of AI tools does matter also. I think we're getting to new age of the "art of mastering AI tools". Like it was the art of pencil-drawing, paint-drawing, digital-illustration.... and it's awsome.

 

[Tom.G]

Yet another opportunity for rolling with laughter:
https://news.mit.edu/2023/researchers-chart-captions-ai-vistext-0630

I can just see a chart labeled
"Number Of Dogs Needed To Walk a Spider"